← Agents / agents / harnesses / think

Client tools

Think supports tools that execute in the browser. The client sends serializable tool schemas in the chat request body, Think merges them with server tools, and when the LLM calls a client tool, the call is routed to the client for execution.

For dynamic client-side tools, pass tools to useAgentChat. Tools with an execute function are registered with the server as client-executed tools:

const { messages, sendMessage } = useAgentChat({
	agent,
	tools: {
		getUserTimezone: {
			description: "Get the user's timezone from their browser",
			parameters: {},
			execute: async () => {
				return Intl.DateTimeFormat().resolvedOptions().timeZone;
			},
		},
		getClipboard: {
			description: "Read text from the user's clipboard",
			parameters: {},
			execute: async () => {
				return navigator.clipboard.readText();
			},
		},
	},
});

const { messages, sendMessage } = useAgentChat({
	agent,
	tools: {
		getUserTimezone: {
			description: "Get the user's timezone from their browser",
			parameters: {},
			execute: async () => {
				return Intl.DateTimeFormat().resolvedOptions().timeZone;
			},
		},
		getClipboard: {
			description: "Read text from the user's clipboard",
			parameters: {},
			execute: async () => {
				return navigator.clipboard.readText();
			},
		},
	},
});

Client tools are tools without an execute function on the server — they only have a schema. When the LLM produces a tool call for one, Think routes it to the client.

For most apps, prefer defining tools on the server and using onToolCall for browser-only execution. The tools option is most useful for SDKs or platforms where the browser decides the available tool surface at runtime.

When a parent agent delegates to a Think sub-agent over RPC with chat() (rather than the browser WebSocket), there is no WebSocket to carry clientTools or to send tool results back. Pass them through ChatOptions instead:

await child.chat(message, callback, {
	signal,
	clientTools: [
		{
			name: "get_user_timezone",
			description: "Get the caller's timezone",
			parameters: { type: "object" },
		},
	],
	onClientToolCall: async ({ toolName, input }) => {
		// Run the client tool wherever the parent can — return its output.
		return runClientTool(toolName, input);
	},
});

await child.chat(message, callback, {
	signal,
	clientTools: [
		{
			name: "get_user_timezone",
			description: "Get the caller's timezone",
			parameters: { type: "object" },
		},
	],
	onClientToolCall: async ({ toolName, input }) => {
		// Run the client tool wherever the parent can — return its output.
		return runClientTool(toolName, input);
	},
});

• clientTools registers the tool schemas for the turn, exactly like the WebSocket clientTools field.
• onClientToolCall executes a client-tool call and returns its output. The model can call a client tool, receive the result, and continue — all within the single chat() call.

If you omit onClientToolCall, the tools are registered but have no result: the model's call is surfaced through the stream callback and the turn ends with a dangling tool call (the RPC stream callback has no inbound result channel of its own). Supply onClientToolCall whenever you want the round trip to complete.

• Recovery: the schemas and onClientToolCall executor are per-turn only and are never persisted (the executor is a live RPC reference that dies with the isolate, and unlike the WebSocket path there is no client to replay a tool-result after an eviction). If an eviction interrupts the turn while a client-tool call is mid-flight, chat recovery errors the orphaned call (treating it like a server tool) and the model proceeds. To re-run cleanly, the parent re-invokes chat() with the clientTools and onClientToolCall again.
• Errors: if onClientToolCall throws, the failure is surfaced to the model as a tool error (output-error) and the turn continues — it does not crash the turn.
• Serialization: the value returned from onClientToolCall becomes the tool output, so it must be JSON-serializable (it travels back over RPC and into the model context).
• No approval gate: RPC client tools execute immediately through onClientToolCall. The WebSocket approval flow (needsApproval) does not apply on this path — gate execution inside your executor if you need it.
• Name precedence: client tools are merged after server tools, so a client tool that shares a name with a server tool (for example a workspace tool) overrides it for that turn — the same as the WebSocket path.
• Abort: aborting the turn via signal stops the loop, but an in-flight onClientToolCall is not itself cancelled; the turn ends after the current call resolves.

Handle browser-side tool execution on the client with onToolCall:

useAgentChat({
	agent,
	onToolCall: async ({ toolCall, addToolOutput }) => {
		if (toolCall.toolName === "read") {
			const result = await readFromBrowser(toolCall.input);
			addToolOutput({
				toolCallId: toolCall.toolCallId,
				output: result,
			});
		}
	},
});

useAgentChat({
	agent,
	onToolCall: async ({ toolCall, addToolOutput }) => {
		if (toolCall.toolName === "read") {
			const result = await readFromBrowser(toolCall.input);
			addToolOutput({
				toolCallId: toolCall.toolCallId,
				output: result,
			});
		}
	},
});

After a client tool result is received, Think automatically continues the conversation without a new user message. The continuation turn has continuation: true in the TurnContext, which you can use in beforeTurn to adjust model or tool selection.

When a turn produces several client tool calls at once, Think waits for all of their results before starting a single continuation, instead of starting one continuation per result. An immediate resume request that arrives while a continuation is already pending attaches to that pending continuation rather than starting a duplicate, and server-side needsApproval continuations resume reliably once the approval is recorded.

A Durable Object can be evicted at any time, including while a turn is paused on an approval prompt or a client-side tool call. Think durable recovery is always on. The SDK treats such a turn as waiting on the human, not stuck. It parks the turn instead of failing it. The user's eventual approval or tool result resumes the conversation.

For which interactions are exempt from recovery budgets, refer to Turns waiting on a human are not sealed.

The messageConcurrency property controls how overlapping user submits behave when a chat turn is already active.

import { Think } from "@cloudflare/think";

export class SearchAgent extends Think {
	messageConcurrency = "latest";
	getModel() {
		/* ... */
	}
}

import { Think } from "@cloudflare/think";
import type { MessageConcurrency } from "@cloudflare/think";

export class SearchAgent extends Think<Env> {
	override messageConcurrency: MessageConcurrency = "latest";
	getModel() {
		/* ... */
	}
}

Think broadcasts streaming responses to all connected WebSocket clients. When multiple browser tabs are connected to the same agent, all tabs see the streamed response in real time. Tool call states (pending, result, approval) are broadcast to all tabs.

Programmatic chat() turns and clearMessages() also broadcast message updates to connected useAgentChat clients, so browser clients stay in sync without reconnecting.